Novel technique pollutes pilfered knowledge graph data

Stolen knowledge graphs incorporated into a GraphRAG AI system could be rendered unusable with the new Active Utility Reduction via Adulteration technique, or AURA, The Register reports.

AURA, which corrupts KG responses to large language models in a bid to produce less accurate predictions without additional pressure to GraphRAG, has shown 100% accuracy in fetching adulterated content and 94% accuracy in producing incorrect responses based on the provided misinformation, according to a study by researchers from the Chinese Academy of Sciences, National University of Singapore, Nanyang Technological University, and Beijing University of Technology. Unlike other poisoned data detoxifiers, AURA has been able to withstand semantic consistency, graph-based, and hybrid checks.

"Fully encrypting the text and embeddings would require decrypting large portions of the graph for every query. This process introduces prohibitive computational overhead and latency, making it unsuitable for real-world use," said researchers.